Closed andresuribe87 closed 1 month ago
Hi, Thank you! This is indeed a typo.
Stop tokens are used for vllm, and here is the detailed description in vllm docs (see stop
): https://docs.vllm.ai/en/v0.5.0/dev/sampling_params.html
The reason for putting so many stop tokens into the generation pipeline is that sometimes the LLM will not generate <|eot_id|> after generating instructions. So we put all special tokens here to have a higher chance of letting it stop :)
This updates the model config to what I believe were typo's.
@fly-dust I'm curious where the definition of the
stop_tokens
comes from? I tried searching online but couldn't figure it out :(